DAP-MR: Degradation-Aware Prompt-Magnitude Routing for Real-World Ultrasound Super-Resolution
Authors: Ngoc-Hoang-Lam Le, Thi-An Le, Dinh-Hung Le, Manh-Hung Nguyen, Kai-Lung Hua, Ching-Chun Huang
Organizations: National Yang Ming Chiao Tung University, Hsinchu, Taiwan · Ho Chi Minh City University of Technology and Education, Ho Chi Minh, Vietnam · National Taiwan University of Science and Technology, Taipei, Taiwan
When super-resolution models designed for natural images are used on ultrasound scans, relying on semantic text prompts can lead to unreliable results. The tags these models generate often do not match the actual restoration needs, and specialized medical terms may be misunderstood by general-purpose text encoders. Instead, prompts focused on common image degradations are more directly useful for guiding restoration. However, using a single prompt to describe both the type and severity of image degradation can entangle these factors, making it hard for the model to know how strong the degradation is. To address this limitation, we introduce Degradation-Aware Prompt--Magnitude Routing (DAP-MR). Each degradation type is represented by a dedicated token initialized from the contextualized EOS embedding of a degradation-oriented prompt and subsequently adapted during ultrasound SR training. A separate magnitude router predicts a continuous, input-dependent magnitude that modulates the corresponding token. Thus, language provides a degradation-aware initialization for what should be restored, while magnitude routing supplies an explicit signal for how strongly it should influence restoration. DAP-MR can be added to both one-step and multi-step diffusion super-resolution frameworks without changing their core structures. Our experiments--on synthetic datasets with both in-domain and under anatomical-domain shift, as well as unpaired real-world ultrasound data--show that DAP-MR delivers better restoration quality and generalizes more consistently across different domains.
Figures & tables
Figure 1. Motivating comparison of prompt strategies using the same PiSA-SR backbone. Natural tags produce limited changes over the no-text baseline, while medical captions may trigger unintended restoration behavior. Degradation prompts provide stronger restoration guidance, but may introduce over-smoothing or artifacts. DAP-MR separates degradation identity from magnitude, producing a result closer to the ground truth. The bottom row shows enlarged regions marked by the green boxes.
Figure 2. Overview of DAP-MR. (a) The magnitude router is pretrained on synthetic LR images with continuous degradation-magnitude labels. (b) Degradation tokens are initialized from contextualized CLIP EOS representations and jointly optimized with the SR backbone. A training mode q∈{full,partial,mismatch} determines the target y(q) and control magnitude m~(q) . Full, partial-dose, and mismatch-specificity supervision learn the clean endpoint, magnitude-dependent restoration, and degradation-selective responses, respectively. The resulting magnitude-modulated degradation tokens condition the original text-guided SR pathway.
Backbone
Variant
Steps
PSNR ↑
SSIM ↑
LPIPS ↓
Cardiac: In-domain
SeeSR
Baseline
50
26.815
0.8888
0.0718
+ DAP-MR
26.723(−0.34%)
0.8901(+0.15%)
0.0748(−4.18%)
StableSR
Baseline
50
26.650
0.8873
0.0813
+ DAP-MR
26.678(+0.11%)
0.8881(+0.09%)
0.0822(−1.11%)
TVT
Baseline
1
26.523
0.9038
0.0724
Table 1. Cross-backbone comparison on the paired synthetic Cardiac and BUSI test sets. Cardiac and BUSI represent in-domain and out-of-domain evaluation, respectively. Percentages shown beside each DAP-MR result indicate relative improvement (positive indicates better performance) from the corresponding baseline.
Backbone
Variant
CAMUS-Poor ↑
HMC-QU ↑
SeeSR
Baseline
75.00
59.07
+ DAP-MR
75.30(+0.40%)
59.66(+1.00%)
StableSR
Baseline
74.27
59.97
+ DAP-MR
74.71(+0.59%)
60.07(+0.17%)
TVT
Baseline
71.70
48.15
+ DAP-MR
74.80(+4.32%)
57.19(+18.77%)
Table 2. Downstream segmentation performance using the frozen SAMUS evaluator. Dice scores are reported in percentage. Values in parentheses indicate relative changes from the corresponding baseline.
Figure 3. Qualitative comparison of baseline and DAP-MR results on the in-domain Cardiac and out-of-domain BUSI test sets across PiSA-SR, TVT, StableSR, and SeeSR. Yellow boxes indicate regions of interest for local comparison.
Full token
EOS token
Token update
Partial
Mismatch
Cardiac
BUSI
CAMUS-P Dice ↑
HMC-QU Dice ↑
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
✓
–
26.394
0.8866
0.0750
22.885
0.6814
0.4291
73.83
52.46
✓
Fixed
26.550
0.8854
0.0777
24.112
0.7223
0.3896
73.84
54.32
✓
Fixed
✓
26.425
0.8842
0.0797
24.391
0.7069
0.3775
74.11
56.10
✓
Fixed
✓
26.426
0.8835
0.0819
24.618
0.7211
0.3462
73.92
53.29
✓
Fixed
✓
✓
26.717
0.8900
0.0769
24.479
0.7119
0.3796
74.26
56.08
Table 3. Ablation of degradation-token representation and supervision strategies using PiSA-SR. Cardiac is in-domain, while BUSI, CAMUS-Poor, and HMC-QU are unseen during training. Partial-only and mismatch-only variants use a 40% sampling ratio. Best and second-best results are shown in bold and underlined.
Label mag.
Router mag.
Cardiac
BUSI
CAMUS-P Dice ↑
HMC-QU Dice ↑
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
✓
26.746
0.8887
0.0892
24.598
0.7130
0.3780
74.10
56.38
✓
26.566
0.8869
0.0759
24.273
0.6825
0.4084
74.13
55.65
✓
✓
26.881
0.8943
0.0842
25.045
0.7616
0.3460
74.26
57.32
Table 4. Ablation of magnitude sources used during full-restoration training. “Label” uses ground-truth degradation magnitudes, “Router” uses router-predicted magnitudes, and enabling both denotes the scheduled transition from label to router magnitudes. Best and second-best results are shown in bold and underlined.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Degradation
Severity parameter
Sampling range
Magnitude normalization
Blur
Blur width σ
σ∼U(0,3.0)
mk=3.0σ
Resize
Downsampling scale s
s∼U(1,4.0)
mk=3.0s−1
Attenuation
Attenuation strength αatt
αatt∼U(0,0.60)
mk=0.60αatt
Speckle
Speckle strength αspk
αspk∼U(0,0.35)
mk=0.35αspk
Sensor noise
Noise std. σn
σn∼U(0,0.06)
mk=0.06σn
Anechoic haze
Haze strength h
h∼U(0,0.12)
mk=0.12h
Appendix
Table 5. Synthetic degradation settings used for online LR generation. For each degradation, pk denotes its severity parameter and mk=ψk(pk)∈[0,1] is the normalized magnitude. The lower and upper bounds correspond to no degradation and maximum severity, respectively.
Setting
Value
Input representation
Frozen VAE latent
Router architecture
MLP
Number of parameters
7,089,030
Output dimension
K=6
Output activation
Sigmoid
Training iterations NR
3000
Appendix
Table 6. Magnitude-router implementation and training settings.
Figure 5. Magnitude-control comparison under blur degradation. Given the same blur-degraded LR input, we vary the requested restoration strength from 0 to 1 using linguistic severity, numerical severity, and the proposed DAP-MR. DAP-MR exhibits the clearest and most progressive response to increasing magnitude, producing increasingly sharper structural details as the control value approaches 1 , whereas text-based severity representations show weaker or less consistent changes.
Figure 6. Router evaluation on the in-domain Cardiac test set and the out-of-domain BUSI set. Importantly, the router is trained only on the in-domain Cardiac training set and is evaluated on BUSI without target-domain adaptation. (a) AUROC measures whether the router correctly detects the presence of each degradation, while active-only MAE evaluates magnitude estimation when that degradation is present. (b) The severity-response curves show how the predicted magnitude changes as the applied degradation severity increases. The router reliably detects degradation presence and generally preserves severity ordering, although its magnitude estimates remain conservative, particularly on the out-of-domain BUSI set.
Prompt strategy
Cardiac
BUSI
CAMUS-P Dice ↑
HMC-QU Dice ↑
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Linguistic
26.800
0.8931
0.0727
23.050
0.6847
0.4093
74.11
55.39
Numerical
26.867
0.8943
0.0878
24.906
0.7469
0.3674
74.09
54.89
RAMTag
26.877
0.8920
0.1013
25.031
0.7611
0.3590
74.24
56.40
DAP-MR (Fixed EOS)
26.717
0.8900
0.0769
24.479
0.7119
0.3796
74.11
54.48
DAP-MR (Learnable EOS)
26.881
0.8943
0.0842
25.045
0.7616
0.3460
74.26
57.32
Appendix
Table 7. Comparison of prompt representations using PiSA-SR. Cardiac is in-domain, while BUSI, CAMUS-P, and HMC-QU are unseen during training. DAP-MR uses degradation-specific EOS embeddings initialized from CLIP and jointly optimized during SR training. Best and second-best results are shown in bold and underlined.
Backbone
Configuration
Params (M)
FLOPs (G)
PiSA-SR
Without router
1289.966
4338.049
With router
1297.055
4338.341
Overhead
+7.089
+0.292
Appendix
Table 8. Parameter and computational overhead of DAP-MR on PiSA-SR at 512×512 resolution. FLOPs include profiled convolution and matrix multiplication operations.
Figure 7. Magnitude sweep for depth-dependent attenuation at fixed LR degradation severity m=0.70 . The router is bypassed, and the attenuation-control magnitude s∈{0,0.25,0.5,0.75,1.0} is manually supplied. The rows show the restored output, zoomed ROI, residual Rs , and signed residual change ΔRs=R0−Rs .
Figure 8. Magnitude sweep for blur at fixed LR degradation severity m=0.70 . The router is bypassed and the blur-control magnitude s∈{0,0.25,0.5,0.75,1.0} is manually supplied. The rows show the restored output, residual Rs , and signed residual change ΔRs=R0−Rs .
Figure 9. Magnitude sweep for resolution degradation at fixed LR degradation severity m=0.70 . The router is bypassed and the corresponding control magnitude is varied from s=0 to s=1 . Positive regions in ΔRs indicate reduced reconstruction residual relative to the zero-control baseline.
Figure 10. Magnitude sweep for speckle noise at fixed LR degradation severity m=0.70 . Increasing s changes only the speckle-specific conditioning magnitude while the remaining degradation controls are kept at zero.
Figure 11. Magnitude sweep for electronic sensor noise at fixed LR degradation severity m=0.70 . The output and residual maps illustrate how the restoration response changes as the noise-specific conditioning magnitude is increased.
Figure 12. Magnitude sweep for anechoic haze at fixed LR degradation severity m=0.70 . The router is bypassed and the haze-control magnitude is manually varied over s∈{0,0.25,0.5,0.75,1.0} .
Medical image super-resolution reconstructs fine spatial detail from low-resolution scans to support clearer anatomy and more useful clinical measurements.
Real-world image super-resolution is particularly challenging for diffusion models because real degradations are complex, heterogeneous, and rarely modeled explicitly. We propose a degradation-aware and structure-preserving diffusion framework for real-world SR. Specifically, we introduce Degradation-aware Token Injection, which encodes lightweight degradation statistics from low-resolution inputs and fuses them with semantic conditioning features, enabling explicit degradation-aware restoration. We further propose Spatially Asymmetric Noise Injection, which modulates diffusion noise with local edge strength to better preserve structural regions during training. Both modules are lightweight add-ons to the adopted diffusion SR framework, requiring only minor modifications to the conditioning pipeline. Experiments on DIV2K and RealSR show that our method delivers competitive no-reference perceptual quality and visually more realistic restoration results than recent baselines, while maintaining a favorable perception--distortion trade-off. Ablations confirm the effectiveness of each module and their complementary gains when combined. The code and model are publicly available at https://github.com/jiyang0315/DASP-SR.git.
Yang Ji, Zonghao Chen, Zhihao Xue +1
KUNBYTE Hangzhou, Zhejiang Province, China · GOLDMYE Xuhui District, Shanghai, China
Recent real-world image super-resolution (SR) methods often adapt text-to-image (T2I) backbones with ControlNet-style branches or spatial conditioning tokens, which increases memory and computes with resolution and often constrains training to a fixed scale. We propose Fill2SR, which repurposes a masked-inpainting Diffusion Transformer for SR without extra spatial branches. Our Inpainting-Interface Evidence Adapter (IIEA) writes the low-quality (LQ) observation into the native masked-image slot under a full-image mask, turning inpainting into a reverse-degradation conditional rectified flow trained with LoRA-only tuning. We further introduce RCDT, an offline pipeline that distills degradation descriptors from unpaired real images and transfers them onto clean targets using frozen open-source models. Fill2SR supports mixed-resolution training up to QHD and yields stable performance across 512/1024/2048 outputs. On synthetic benchmarks, our base model with IIEA achieves the best LPIPS on DIV2K and LSDIR; adding RCDT trades a small LPIPS drop for consistently stronger no-reference quality on RealLQ250 and RealPhoto60. Fill2SR remains memory-predictable, running 15362 inference on a single 32GB GPU and extending to multi-megapixel outputs via tiled restoration.
Xingfu Yi, Xiaoxue Yu
Independent Researcher, Hangzhou, China · Zhejiang University, Hangzhou, China
Low-cost handheld ultrasound devices can be widely deployed compared to professional hospital ultrasound machines. However, their images suffer from compound degradation that can mislead clinical judgment. Motivated by this observation, mapping handheld low-quality (LQ) to hospital high-quality (HQ) images has been considered a valuable research question. Conventionally, the mapping requires pixel-aligned LQ-HQ pairs. This requirement is unsatisfactory in practical scenarios because real scans at different times are never pixel-aligned. This paper addresses the challenge with a two-stage framework. The first stage generates pixel-aligned LQ-HQ datasets, and the second stage trains an enhancement model that improves LQ images. The first stage trains a cycle-consistent style-transfer model on unaligned real LQ-HQ pairs to learn a HQ-to-LQ model. Then, the model transforms real HQ images into pixel-aligned LQ images. Based on the dataset generated by the first stage, the second stage uses the Dual Degradation-Guided (DDG) Low-Rank Adaptation (LoRA) method to fine-tune an LQ-to-HQ model based on aligned pairs. In this stage, the model is based on the well known PiSA-SR framework but inserts a degradation-conditioned correction matrix. Experimental results on the USenhance2023 dataset show that the FID metric is improved by 16.7% over the strongest baseline while other metrics indicate that our enhanced outputs are well aligned with the real HQ distribution. The source code of our method is available at https://github.com/Jason0411202/DDG_LoRA.
Yu-Kai Wang, Chun-Xin Tan, Manh-Hung Nguyen +1
Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan · Ho Chi Minh City University of Technology and Engineering, Ho Chi Minh City, Vietnam